Research map / Probabilistic and scientific modeling / Dynamical systems
THD-BAR: Topology Hierarchical Derived Brain Autoregressive Modeling for EEG Generic Representations
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes THD-BAR, an EEG foundation model that imposes a physiologically-grounded multi-scale channel hierarchy (Brain Topology Hierarchy), tokenizes EEG with a multi-scale VQ-VAE, and pre-trains a GPT-2-style autoregressive model with "next-scale-time" prediction.
Problem
Existing EEG autoregressive foundation models order multi-channel EEG purely temporally ("next-time prediction"), which fails to capture dynamic spatial topology of brain activity and limits universal EEG representation learning.
Approach
Define a multi-scale spatial order over EEG channels (whole brain -> brain region -> channel), tokenize EEG with a topology-hierarchical VQ-VAE producing per-scale token maps, and autoregressively predict tokens scale-by-scale within each time step and time-by-time across steps using a scale-time mask.
Key results
- THD-BAR-Base improves over NeuroLM by 2.2% on DEAP and 3.3% on SEED balanced accuracy.
- THD-BAR-Huge reaches 65.3% balanced accuracy on TUEV vs 45.6% for NeuroLM and 64.1% for LaBraM.
- THVQ-VAE improves PCC by over 8% relative to the VQ-VAE used in NeuroLM during tokenizer training.
- Ablation: the five-scale configuration (O5) gives the best tokenizer and BAR performance; the scale-time mask outperforms scale-only and time-only masks.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use THD-BAR-Base instead of NeuroLM | THD-BAR · DEAP | balanced accuracy: 2.2% | Helped |
| Use THD-BAR-Base instead of NeuroLM | THD-BAR · SEED | balanced accuracy: 3.3% | Helped |
| Scale the model from Base to Huge | THD-BAR · downstream EEG datasets | balanced accuracy | Mixed |
| Use THD-BAR rather than EEGPT | STEW · STEW | balanced accuracy: 63.2$\pm$10.6 vs 62.9$\pm$7.9 | Hurt |
| Use all BTH scales S1–S5 rather than only a single scale (ablation) | THD-BAR · evaluated scale configurations | performance | Helped |
| Use only the coarsest scale S1 (ablation) | THD-BAR · evaluated scale configurations | performance | Hurt |
| Use only the finest scale S5 (ablation) | THD-BAR · evaluated scale configurations | performance | Hurt |
| Combine scale-wise and time-wise masking in the scale-time-wise mask (ablation) | BAR · downstream EEG tasks | accuracy | Helped |
| Use THVQ-VAE instead of VQ-VAE (ablation) | EEG tokenizer · tokenizer training | PCC: over 8% | Helped |
Surprising findings
- EEGPT outperforms THD-BAR on STEW, which the authors attribute to EEGPT being pre-trained and fine-tuned on STEW itself.
Limitations
- Multi-modal integration is limited to instruction-based fine-tuning; deeper fusion with other physiological signals is unexplored.
- Current BTH is a single physiologically-derived partition; alternative topological partitions are not studied.
- Evaluated only on classification tasks; efficiency for real-time BCI is not addressed.
Experimental setup
Datasets: SEED-IV SEED-V SEED-GER SEED-FRA EmoBrain Grasp and Lift EEG Challenge Inria BCI Challenge EEG Motor Movement/Imagery Dataset TUAR TUEP TUSZ TUSL
Benchmarks: DEAP SEED MIBCI BCIC4-1 EEGMat STEW EDF HMC TUAB TUEV
Metrics: balanced accuracy PCC loss perplexity
Baselines: EEGNet TSception LGGNet BIOT LaBraM EEGPT NeuroLM
Builds on: VQ-VAE VAR NeuroLM LaBraM
Closest papers by meaning
- KAST-BAR: Knowledge-Anchored Semantically-Dynamic Topology Brain Autoregressive Modeling for Universal Neural Interpretation (ICML 2026 · similarity 0.90)
- Uni-NTFM: A Unified Foundation Model for EEG Signal Representation Learning (ICLR 2026 · similarity 0.88)
- CSBrain: A Cross-scale Spatiotemporal Brain Foundation Model for EEG Decoding (NeurIPS 2025 · similarity 0.85)
- BaRISTA: Brain Scale Informed Spatiotemporal Representation of Human Intracranial Neural Activity (NeurIPS 2025 · similarity 0.85)
- B[FM]$^2$: Brain Foundation Model via Flow Matching with SplitUNet (NeurIPS 2026 · similarity 0.84)
- LUNA: Efficient and Topology-Agnostic Foundation Model for EEG Signal Analysis (NeurIPS 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.